Paper orchestra

paper-orchestra adapter for EZ Math Model

This is the EZMM-facing adapter for the vendored upstream PaperOrchestra pack at external/paper-orchestra/. It lets the EZMM writer call PaperOrchestra's writing skills without changing EZMM's contest-paper contract.

Upstream source: https://github.com/Ar9av/PaperOrchestra

Paper citation: Song et al., "PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing", arXiv:2604.05018, 2026.

When writer calls this

Load this adapter in pipeline/04-paper-writing before drafting paper.md. Use it to create a structured writing workspace, generate or refresh a paper outline, organize literature-review tasks, draft the body, and run a refinement pass.

Do not use this adapter to bypass EZMM hard rules. The final paper.md must still satisfy prompts/writer.md, references/roles/writer-guide.md, and pipeline/05-quality-audit.md.

Source roots

All paths below are relative to skills/ez-math-model/.

Purpose	Path
EZMM adapter	tools/paper-orchestra/SKILL.md
Vendored upstream pack	external/paper-orchestra/
Upstream orchestrator	external/paper-orchestra/skills/paper-orchestra/SKILL.md
Upstream outline agent	external/paper-orchestra/skills/outline-agent/SKILL.md
Upstream literature agent	external/paper-orchestra/skills/literature-review-agent/SKILL.md
Upstream section writer	external/paper-orchestra/skills/section-writing-agent/SKILL.md
Upstream refinement agent	external/paper-orchestra/skills/content-refinement-agent/SKILL.md
Upstream shared writing checks	external/paper-orchestra/skills/shared/writing_quality_check.md

When running upstream helper scripts, set the working directory to external/paper-orchestra/ so upstream examples like python skills/paper-orchestra/scripts/validate_inputs.py ... resolve correctly.

Default mode: PaperOrchestra-first writing

The writer runs PaperOrchestra-first by default. Do not wait for a separate PaperOrchestra request. Build the PaperOrchestra workspace automatically from EZMM runtime artifacts, run the upstream outline, literature, section-writing, and refinement protocol as far as the local inputs and tools allow, then bridge the result back to EZMM's required runtime/{task_id}/paper.md.

When a usable LaTeX template and TeX toolchain are available or can be safely scaffolded, also produce the upstream-style package under runtime/{task_id}/paper_orchestra/final/. The EZMM-compatible paper.md remains mandatory because downstream quality audit and packaging consume it.

Create a local workspace:

runtime/{task_id}/paper_orchestra/
├── inputs/
│   ├── idea.md
│   ├── experimental_log.md
│   ├── conference_guidelines.md
│   └── figures/
├── outline.json
├── refs/
├── drafts/
│   ├── paper.tex
│   └── paper.md
├── refinement/
│   └── worklog.json
├── final/
│   ├── paper.tex
│   └── paper.pdf
└── adapter_report.md

Map EZMM inputs into PaperOrchestra-style inputs:

PaperOrchestra input	EZMM source
idea.md	problem.md, intake.json, modeling_plan.md, and the intended contribution of each quesN
experimental_log.md	execution_log.md, src/*.py final print blocks, `results/*.csv
conference_guidelines.md	contest name, language, run_mode, section requirements from templates/chapter_outline.toml, plus EZMM no-bullet and no-leakage rules
inputs/figures/	copies or references to figures/*.png with usable_in_paper=true

The adapter report must record which EZMM files were consumed, which accepted figures must appear in the paper, and whether any PaperOrchestra sub-step was degraded or skipped.

Writing workflow

1. Read upstream outline-agent/SKILL.md and use its schema discipline to produce paper_orchestra/outline.json. For EZMM, section_plan must map to the contest sections in templates/chapter_outline.toml.
2. Read upstream literature-review-agent/SKILL.md for search, verification, deduplication, and citation-pool discipline. Prefer EZMM tools/scholar/SKILL.md or tools/paper_search/SKILL.md for actual metadata retrieval, because those tools already follow the EZMM env and setup policy.
3. Read upstream section-writing-agent/SKILL.md for single-pass global coherence, table construction, figure-text alignment, and citation-key discipline. Prefer the upstream LaTeX draft path when the workspace has a usable template; then convert or mirror the accepted content into EZMM Markdown. If LaTeX inputs or tools are unavailable, adapt LaTeX-only instructions directly to Markdown:
   • figures stay as ![描述](figures/name.png);
   • tables stay as Markdown tables unless paper_*.md template requires otherwise;
   • formulas stay in pandoc-compatible $...$ / $$...$$;
   • do not introduce LaTeX preamble, \section, \cite, or BibTeX in paper.md.
4. Read upstream content-refinement-agent/SKILL.md and run refinement by default unless the draft is blocked by missing factual inputs. Accept a revision only when it improves correctness, figure-text binding, citation integration, or clarity without weakening EZMM constraints.
5. Write the final Markdown to runtime/{task_id}/paper.md only after the adapter's self-check passes.

Upstream LaTeX package

Run the upstream orchestrator automatically when the required workspace/inputs/template.tex and workspace/inputs/conference_guidelines.md can be supplied from EZMM templates or local contest metadata. Use external/paper-orchestra/skills/paper-orchestra/SKILL.md unchanged for that sub-run and write outputs under paper_orchestra/final/.

If the full LaTeX route is not runnable, do not ask the user. Record the reason in adapter_report.md, continue with the Markdown adaptation path, and keep the same PaperOrchestra outline/literature/section/refinement discipline.

Required self-check before returning to writer

Before this adapter hands control back to the writer stage, confirm:

• paper_orchestra/outline.json exists or adapter_report.md explains why it was skipped.
• adapter_report.md records whether the full upstream LaTeX route ran or which missing input/tool caused an automatic Markdown adaptation.
• Every chart_manifest.json entry with status=accepted and usable_in_paper=true is listed in adapter_report.md.
• paper.md references only figures allowed by EZMM chart manifest.
• No PaperOrchestra workspace paths, upstream script paths, runtime/, adapter_report.md, or outline.json leak into the paper body.
• Citations are unique and compatible with EZMM's {[^N] ...} protocol.
• Any skipped PaperOrchestra sub-step is recorded as a quality-audit note.

Licensing note

The vendored upstream pack is MIT licensed. Keep the upstream LICENSE, CITATION.cff, and README under external/paper-orchestra/, and cite PaperOrchestra when its workflow materially shapes the final paper.